Development of Neural Network Approach for Reconstructing the Atmospheric Humidity Profile from Ground-based Microwave Radiometer-spectrometer Data

An original neural network approach based on the mixture of experts has been developed to solve the inverse problem of absolute humidity altitude profile retrieval from atmospheric downwelling radiation brightness temperature obtained in the bands of water vapor resonant absorption near 22.235 and 183.3GHz lines. The software implementation of the neural network model assumes the possibility of its training on GPU. The training dataset has been formed by repeatedly solving the forward problem of simulating the brightness temperature spectra based on around 3000 sessions of radiosonde measurements of meteorological parameters profiles (up to an altitude of 15km) conducted in the city of Dolgoprudny, Moscow Region, from January 2019 to October 2023. These meteorological parameters include air temperature T, atmospheric pressure P, and relative humidity, which is than converted to absolute humidity values ρ. For the dataset augmentation purposes, the temperature profile T varies within±5%, and absolute humidity profile ρ within±20% from the corresponding near-surface values. Herewith, the horizontally homogeneous stratus cloud layer, which vertical extent gradually increases while its liquid water content also increases from 0kg/m
²
to more than 5.5kg/m
²
, has been modeled in the field-of-view of ground-based microwave radiometer antenna. The results of applying the trained model to the test dataset show satisfactory agreement between the reconstructed absolute humidity profiles and the values measured by the radiosonde.